Compare nine voice agent APIs on price, latency claims, telephony, and coding-agent tooling, using each vendor's own published figures.
The short answer
The right voice agent API depends on how much of the stack you want to own. If you want to pick your own speech-to-text, LLM, and voice behind a hosted orchestrator, the usual choices are Vapi ($0.05 per minute plus providers at cost) and Retell ($0.07 to $0.31 per minute all-in). To own and self-host the code, LiveKit Agents is the reference framework. For one model that hears, reasons and calls tools in a single session, OpenAI Realtime is the default, and ElevenLabs ($0.08 per minute) and Deepgram ($0.075) sell managed versions. Telnyx ($0.05 per minute plus tokens and carrier) and Bland ($0.12 to $0.14, telephony included) put the carrier and the agent on one bill.
PolyAI sells two products, for two different jobs. Dialog-RSN-1 is a real-time model at 2 cents per connected minute plus tokens, billed per second. It speaks the OpenAI Realtime protocol, so the stock OpenAI SDK connects to it, and it can return text for your own TTS voice. It is English only for now. Dialog Studio is for enterprise customer service, where you need phone numbers, contact center handoff, and analytics in one place.
Comparison table
All prices in USD, list price, pay-as-you-go tier. "Latency" is the vendor's own published claim, quoted as written; vendors measure different things, so the numbers are not comparable with each other. We ran no latency tests.
| Vendor / product | Pricing unit | Latency claim as published | Languages | Telephony | Self-serve | MCP / coding-agent tooling | Best for |
|---|---|---|---|---|---|---|---|
| PolyAI Dialog Studio | Per minute; rate not published ("talk to our team") | None published for the platform | 73 response languages | PolyAI numbers, your Twilio, self-serve custom SIP, Amazon Connect, Five9 | Yes, studio.poly.ai | Builder MCP and Data MCP; agentic setup prompt for Claude Code, Cursor, Codex | Enterprise customer-service agents needing telephony, analytics and contact center handoff in one platform |
| PolyAI Dialog-RSN-1 API | 2 cents per connected minute plus tokens, billed per second | "Sub-300ms median model first token" | English only (more announced) | Not included; bring your own | Yes | Dialog-RSN-1 skill for Claude Code, Codex, Cursor, Gemini CLI; docs MCP | Developers with their own stack who want a realtime model on the OpenAI Realtime protocol |
| Vapi | $0.05 per minute hosting, plus STT, LLM, TTS at provider cost | "Sub-500ms latency" / "<500ms average latency" | Provider-dependent | Vapi numbers and SIP, Twilio, Vonage, Telnyx, BYO SIP | Yes, $5 credit | CLI, agent skills, docs MCP | Mixing providers behind a hosted orchestrator |
| Retell AI | $0.07 to $0.31 per minute all-in | "~600ms latency" | Provider-dependent | Retell numbers, Twilio, Telnyx, Vonage, SIP, Genesys, Five9, Avaya | Yes, $10 credit, 20 concurrent | Hosted MCP, CLI, SDKs | Dashboard-first builder with an all-in price |
| LiveKit (Agents + Cloud) | Open source; Cloud sessions $0.01 per minute after 1,000 included | "under 250ms" (media transport, not agent response) | Provider-dependent | SIP in and out; 1 free number | Yes | Docs MCP, CLI | Owning the agent code, self-hosting |
| OpenAI Realtime / GPT-Live | gpt-realtime-2.1: $32 / $64 per 1M audio tokens. GPT-Live 1: $0.05 per minute plus backend model | None on pages checked | Not stated [unverified] | Direct SIP; partner guides | Yes | Function and MCP tools in session; Agents SDK | One-model speech-to-speech; teams on OpenAI |
| ElevenLabs Agents | $0.08 per minute; LLM passed through | 620ms (Mini), 570ms (Max) time to first audio | 90+ | Twilio, SIP, Telnyx, Plivo, Vonage | Yes | Hosted MCP and plugin | Voice quality and choice |
| Deepgram Voice Agent API | $0.075 per minute; $0.050 with own LLM and TTS | Flux end of turn "about 260 ms" | 60+ (STT) | Guides for Twilio, Plivo, Amazon Connect | Yes, $200 credit | CLI as MCP, agent skills | Flat rate from a speech-recognition vendor |
| Bland | $0.14 per minute (Start); $0.12 (Build, $299/month); all included | "Sub-400ms response latency" | 40+ | Included; BYO SIP | Yes | MCP server and plugin | One per-minute line item |
| Telnyx Voice AI | $0.05 per minute plus LLM tokens and carrier from $0.0032 per minute | "Sub-200ms response times" | 80+ | Telnyx is the carrier | Yes | Remote MCP, agent skills | Cost-sensitive phone traffic at volume |
Vendor by vendor
PolyAI
PolyAI sells two developer products. Dialog Studio is a self-serve platform (sign-up at studio.poly.ai) for building, testing, and running voice and chat agents, with three REST API families (Agents, Conversations, Webhooks), hosts in the US, UK, and EU, and contact center integrations including custom SIP, Twilio, Amazon Connect, and Five9. Developers can build and query agents from an IDE through two hosted MCP servers: Builder MCP (create, test and deploy agents) and Data MCP (search conversations, read transcripts, aggregate metrics). Platform usage is priced per minute, but the rate is not published; you have to contact sales. PolyAI's API reference states 99.994% actual uptime over the trailing 12 months (measured, not the SLA).
The Dialog-RSN-1 API is a real-time model over WebSocket that hears audio, decides when the caller has finished, and writes the reply; by default, the server also speaks it, or you can take text and use your own TTS. It costs 2 cents per connected minute plus tokens, billed per second. It speaks the OpenAI Realtime protocol, and the docs include a migration guide from OpenAI Realtime and a LiveKit Agents guide. The docs ship a quickstart for coding agents: install the Dialog-RSN-1 skill in Claude Code, Codex, Cursor, or Gemini CLI and build a working voice agent in three prompts. Limits: English only today, 2 concurrent sessions on the sign-up credit and 10 once a card is on file, 60-minute sessions, 32,768-token context, no telephony included. PolyAI's pricing page states "sub-300ms median model first token".
Choose something else if you need Dialog-RSN-1 in a language other than English, or more than 10 concurrent sessions without talking to sales.
Vapi
Vapi is a hosted orchestrator. You pick a transcriber, LLM, and voice from its catalog; Vapi charges $0.05 per minute for hosting and passes provider costs through without markup. The free tier includes $5 of credit, 4 concurrent calls, and one phone number; paid packages start at $29 per month (10 concurrent calls) and $999 per month (30). HIPAA is a $2,000 per month add-on. Fits better than PolyAI when you want to swap providers per call and see each cost line.
Retell AI
Retell prices voice agents at $0.07 to $0.31 per minute all-in, set by your LLM and voice choice. New accounts get $10 of credit and 20 concurrent calls. It has a hosted MCP server, a CLI and SDKs, and the widest list of documented contact center SIP guides among the orchestrators here (Genesys, Five9, Avaya). Retell publishes "~600ms latency". Fits better than PolyAI for teams that want a published, all-in self-serve price with no sales step.
LiveKit
LiveKit Agents is an open-source framework for Python and Node.js, built on LiveKit's open-source WebRTC server; you can self-host or use LiveKit Cloud. Cloud's Build plan is $0 per month with 1,000 agent session minutes, then $0.01 per minute, with model inference and telephony priced separately. Dialog-RSN-1, OpenAI, and Gemini realtime models all run inside LiveKit Agents. Fits better than PolyAI when you want to own and modify the whole agent runtime.
OpenAI Realtime API and GPT-Live
The Realtime API (gpt-realtime-2.1 at $32 / $64 per 1M audio tokens in / out, gpt-realtime-2.1-mini at $10 / $20) runs speech, reasoning, and tool use in one session. GPT-Live 1 is a full-duplex conversation model at $0.05 per minute that delegates reasoning and tools to a separate backend model, billed separately. Phone calls can connect over direct SIP. Fits better than PolyAI when you want the Realtime protocol from its originator, or need languages Dialog-RSN-1 does not yet cover.
ElevenLabs Agents
ElevenLabs Agents charges $0.08 per call minute on its plan tiers, with LLM costs passed through, and a 95% discount on silences longer than 10 seconds. Its Realtime v1 Mini and Max models are speech-to-speech, OpenAI Realtime-compatible, support 90+ languages, and publish 620ms and 570ms time to first audio; the docs say they are rolling out gradually. Fits better than PolyAI when the voice itself (cloning, a large voice library) is the priority.
Deepgram Voice Agent API
Deepgram runs speech-to-text, a managed LLM and text-to-speech over one WebSocket for $0.075 per minute, down to $0.050 with your own LLM and TTS. New accounts get $200 of credit with no card. Telephony is bring-your-own. Fits better than PolyAI when you want a flat per-minute rate from a speech-recognition specialist and plan to handle telephony yourself.
Bland
Bland bundles telephony, STT, LLM and TTS into one rate: $0.14 per minute on the free Start plan (10 concurrent calls, 100 calls per day) and $0.12 on Build ($299 per month, 50 concurrent). It states sub-400ms response latency and 40+ languages. Fits better than PolyAI for high-volume outbound calling with a single bill.
Telnyx Voice AI
Telnyx is a carrier that also runs the agent: $0.05 per minute for orchestration, STT and TTS, plus LLM tokens and carrier cost (inbound SIP from $0.0032 per minute). Voice-engine charges round to 60-second increments per call, which matters for short calls. Fits better than PolyAI when cost per minute at volume and owning the phone network decide.
FAQ
Which voice agent APIs work with Claude Code, Cursor or Codex?
All nine publish something for coding agents. PolyAI (Builder and Data MCP, Dialog-RSN-1 skill), Retell, ElevenLabs, Bland and Telnyx run hosted MCP servers that act on your account; Vapi and Deepgram publish agent skills and a CLI; LiveKit publishes a docs MCP server. OpenAI Realtime supports MCP tools inside a voice session, which is a different thing.
Do I need a separate telephony provider?
With Bland and Telnyx, no. Vapi, Retell, PolyAI Dialog Studio, ElevenLabs and LiveKit sell or connect numbers and accept your own SIP trunk. OpenAI Realtime supports direct SIP. Deepgram and PolyAI Dialog-RSN-1 expect you to bring telephony.
Can I move an OpenAI Realtime app to another provider?
Yes, for providers that implement the same protocol. Dialog-RSN-1 and ElevenLabs Realtime both document compatibility, and PolyAI publishes a migration guide. Not every parameter carries over; Dialog-RSN-1 rejects the reasoning session field.
Which voice agent API is best for enterprise contact centers?
Look for documented connectors to your contact center platform, data residency and a security certification. Retell and PolyAI document Genesys and Five9 integrations; PolyAI also documents Amazon Connect and runs US, UK and EU regions.
How we compiled this
PolyAI wrote this page and is one of the vendors compared. Every price, limit, and latency figure was read from the vendor's own pricing page or documentation on 4 Nov 2026. Latency figures are quoted as published; we ran no benchmarks and do not rank on latency. We used list prices on the entry pay-as-you-go tier. Where we could not confirm a fact on a primary source, we left it out or marked it [unverified]. Vendors who believe a figure is out of date can contact PolyAI and we will correct it.
Sources
- PolyAI: pricing, Dialog-RSN-1 limits, overview, coding-agent quickstart, MCP, API reference and uptime, languages
- Vapi: pricing, docs
- Retell: pricing, docs
- LiveKit: pricing, docs
- OpenAI: pricing, SIP, GPT-Live
- ElevenLabs: agents pricing, API pricing, realtime models
- Deepgram: pricing, Voice Agent API
- Bland: pricing, docs
- Telnyx: pricing, product